A Hybrid Framework to Quantitatively Assess Deformable Image Registration Accuracy with a Complete Implementation of TG-132 Based Validation
Presentations
PO-GePV-M-183 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To develop a framework to analyze deformable image registration (DIR) and quantitatively describe the errors using a manually identified ground truth and image similarity metrics, including those described in TG-132.
Methods: The hybrid nature of the framework stems from the use of a manually identified ground truth and calculated image similarity. The ground truth is identified by manually selecting landmark voxel locations in the reference image and target image. The Euclidean distance between the deformed landmark voxel location following the registration-generated deformation vector field (DVF) and the manually identified target landmark voxel location informs the target registration error (TRE) as described in TG-132. Warp images (target image deformed to reference image geometry) are employed to calculate the dice similarity coefficient (DSC) and mean distance to agreement (MDA) of contours generated from the volumes surrounding the selected landmark. Additionally, the framework employs similarity measures outside of TG-132, such as correlation coefficient and structural similarity.To validate the framework’s usability, we chose free breathing CT scans. We used 23 scan pairs, deformably registered using dense displacement sampling (deeds). 20-50 landmarks were manually selected on both reference and target scans, for each patient. With this input, the framework investigated the DIR accuracy and reported the results of the image similarity metric calculations for each landmark.
Results: 671 landmarks were manually identified. The mean TRE was 1.48 ± 2.19 mm. The mean DSC was 0.97 ± 0.08. The mean MDA was 0.15 ± 0.88 mm. The mean consistency was 2.30 ± 2.79 mm. The mean structural similarity was 0.78 ± 0.11. The mean correlation coefficient was 0.91 ± 0.08.
Conclusion: The framework successfully completed a quantitative assessment of DIR accuracy as well as conducted a full implementation of the TG-132 image registration validation procedure.
Keywords
Taxonomy
Contact Email